VORPS Y12, Y6, K7, Y6                              // 62d14c2f56f4
	VORPS -7(CX)(DX*1), Y6, K7, Y6                     // 62f14c2f56b411f9ffffff
	VORPS -15(R14)(R15*4), Y6, K7, Y6                  // 62914c2f56b4bef1ffffff
	VORPS Y11, Y26, K7, Y6                             // 62d12c2756f3
	VORPS Y26, Y26, K7, Y6                             // 62912c2756f2

	VPBLENDMW (DX), X3, K4, X8                         // 6272e50c6602
	VPBLENDMW Y3, Y0, K2, Y6                           // 62f2fd2a66f3
	VPBLENDMW (R14), Y0, K2, Y6                        // 62d2fd2a6636
	VPBLENDMW -7(DI)(R8*8), Y0, K2, Y6                 // 62b2fd2a66b4c7f9ffffff
	VPBLENDMW Z15, Z3, K2, Z14                         // 6252e54a66f7

#1����}ݾ�^%�v������]}΅`�U��Л��z 5���M�qMV0$}��d�x�Z�����S;�`_�5�nb����~bkM�����܂�W�n31���⌯���ܷvWAU��DE�I#m{���I�:����U(WOU�`���7f��~%# :����s��F��/XIk���ʤ&}�qfu{;�̪���%/�vuN�f�w���N&�9���?�X��:{Li7�u������qy�;+&_%���]i�i[�*h�G��y6�n�H��=�ʻO�V���&O���0ωRs�8\X,{n�;�Uvp��x[���]���OM#}Q�� 3��(6�;1 ���[��o|7�g�'&"���
���� �V�����/�6{e-ܭ;{Z�t������w]��h7��>8�;��$�1yyԹ���-�4�� F�fz#ᬈtٽ��w���������H�������?i4n�z� ���"�/���j�[b���u�ғˀ镪�(�K�}LS����jqU�Ƶ���,�}Z����.e�...

#define ctr0StoreAVX2 (3*32)(BP)
#define ctr1StoreAVX2 (4*32)(BP)
#define ctr2StoreAVX2 (5*32)(BP)
#define ctr3StoreAVX2 (6*32)(BP)
#define tmpStoreAVX2 (7*32)(BP) // 256 bytes on stack
#define AA0 Y0
#define AA1 Y5
#define AA2 Y6
#define AA3 Y7
#define BB0 Y14
#define BB1 Y9
#define BB2 Y10
#define BB3 Y11
#define CC0 Y12
#define CC1 Y13
#define CC2 Y8
#define CC3 Y15
#define DD0 Y4
#define DD1 Y1
#define DD2 Y2